TinyMyo: a Tiny Foundation Model for Flexible EMG Signal Processing at the Edge
Organizations: Integrated Systems Laboratory of ETH Zürich, Zürich, Switzerland · Department of Electrical, Electronic and Information Engineering (DEI), University of Bologna, Bologna, Italy
Abstract
Objective: Surface electromyography (EMG) is widely used in biomechanics, rehabilitation, prosthetic control, and human-machine interfaces. Generalization across subjects, recording systems, and acquisition protocols remains challenging. Existing EMG foundation models (FMs) have limited support for multiple downstream tasks and deployment on low-power devices. Methods: We introduce TinyMyo, a 3.6-million-parameter Transformer encoder pre-trained on public EMG datasets using self-supervised masked reconstruction. TinyMyo supports multiple tasks with minimal changes to the task-specific prediction head. Results: We evaluated TinyMyo on hand gesture classification, hand kinematic regression, speech production, and speech recognition. For gesture classification, TinyMyo outperformed previously reported EMG FMs, achieving accuracies of 87.98% on NinaPro DB5, 97.1% on UCI-EMG, and 96.6% on EPN-612. On the GAP9 ultra-low-power microcontroller, TinyMyo achieved an inference time of 0.785 s and consumed 44.91 mJ per inference at an average power of 57.18 mW. A 1.9-million-parameter variant processed 5 s and 1 s input windows in 0.496 s and 0.089 s, respectively. Conclusion: TinyMyo demonstrates that a compact, self-supervised EMG encoder can support multiple downstream tasks and run on a low-power microcontroller. Significance: To our knowledge, this is the first deployment of an EMG FM on an ultra-low-power microcontroller, opening a path to on-device motor intent decoding, neuromuscular assessment, and human-machine interaction.
Figures & tables
| Dataset | #S | Hz | Ch | Loc. | Metric |
| Hand Gesture Classification | |||||
| Ninapro DB5 [14] | 10 | 200 | 16 | FA | Acc, F1 |
| EPN-612 [15] | 612 | 200 | 8 | FA | Acc, F1 |
| UCI-EMG [16] | 36 | 200 | 8 | FA | Acc, F1 |
| GNI [5] | 100 | 2k | 16 | H/A | CLER |
| Hand Kinematic Regression | |||||
| Dataset | Time | Freq. | ||
|---|---|---|---|---|
| RMSE | RMSE | |||
| Ninapro DB6 [ 20 ] | 0.804 | 0.992 | ||
| Ninapro DB7 [ 33 ] | 0.813 | 0.993 | ||
| EMG2Pose [ 11 ] | 0.809 | 0.990 | ||
| Model | Params | NinaPro DB5 | EPN-612 | UCI-EMG | ||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Acc | F1 | AUROC | Acc | F1 | AUROC | Acc | F1 | AUROC | ||
| Moment (as in [ 12 ] ) | 385M | 86.41 | 74.42 | N.A. | 90.87 | 90.16 | N.A. | 90.45 | 91.75 | N.A. |
| Moment (as in [ 13 ] ) | 86.41 1.13 | 76.84 1.95 | 99.34 0.12 | 93.83 0.76 | 93.75 0.82 | 99.63 0.09 | 92.79 0.64 | 91.24 0.71 | 99.47 0.11 | |
| PhysioWave-Small b [ 12 ] | 5M | 84.78 | 72.54 | N.A. | 93.12 | 93.40 | N.A. | 90.35 | 89.51 | N.A. |
| WaveFormer a, b [ 13 ] | 3.1M | 87.53 0.92 | 74.66 1.78 | 99.35 0.10 | 95.21 0.68 | 95.22 0.71 | 99.70 0.07 | 93.10 0.58 | 93.20 0.62 | 99.60 0.08 |
| TinyMyo (FS), 1 sec | 3.6M | 83.12 0.58 | 61.72 1.63 | 97.57 0.3 | 72.97 0.1 | 72.97 0.14 | 93.38 0.11 | 93.13 0.44 | 93.13 0.46 | 99.57 0.05 |
| Model | Params | CLER | Inference Method |
|---|---|---|---|
| LSTM, Kaifosh [ 5 ] (a) | 6.4M | 0.1819 | Full sequence |
| LSTM, Kaifosh [ 5 ] (b) | 6.4M | 0.1596 | Windowed |
| TinyMyo - FS | 3.6M | 0.142 0.007 | Windowed |
| TinyMyo - FT | 3.6M | 0.146 0.005 | Windowed |
| TinyissimoMyo - FS | 1.8M | 0.151 0.008 | Windowed |
| TinyissimoMyo - FT | 1.8M | 0.153 0.01 | Windowed |
| Model | Params | MAE ∘ | RMSE ∘ |
|---|---|---|---|
| TEMPONet TCN, Zanghieri [ 48 ] ∗ | 500K | 6.89 | – |
| Event-based Lin. Reg., Zanghieri [ 49 ] ∗ | – | 8.8 2.3 | – |
| DNKF, Bao [ 50 ] ∗∗ | – | – | 13.5 |
| TinyMyo - FS, 100 ms ( ∗∗∗ ) | 3.8M | 9.48 0.29 | 15.48 0.46 |
| TinyMyo - FT, 100 ms | 3.8M | 9.05 0.11 | 14.73 0.17 |
| TinyMyo - FS, 500 ms | 3.8M | 9.35 0.26 | 15.18 0.42 |
| Transduction Model | Vocoder | ASR | WER | |||
| Method | Model | Params [M] | Model | Params [M] | ||
| Transformer, Gaddy [ 37 ] | CNN + Transformer | 54 | HiFi-GAN | 13.92 | DeepSpeech | 36% |
| Diff-ETS, Ren [ 51 ] | as [ 37 ] + Diffusion Probabilistic Model | 54 ( a ) | HiFi-GAN | 13.92 | Whisper Medium (EN) | 32% |
| SU-E2S, Scheck. [ 52 ] | HuBERT-Soft | N.A. | HiFi-GAN | 13.92 | Whisper Medium (EN) | 40% |
| DiffMV-ETS [ 53 ] | DiffMV-ETS | 66 (b) | BigVGAN | 112 | Distil-Whisper-v3 | N.A. |
| EMGVox-GAN [ 54 ] | as [ 37 ] | as [ 37 ] | EMGVox-GAN | 12 | N.A. | 36% |
| Method | Params | WER |
|---|---|---|
| Transformer, Gaddy [ 41 ] | 54 M | 28.8% |
| TinyMyo - FT | 4.5 M | 33.95% 0.97% |
| TinyissimoMyo - FT | 2.6 M | 43.07% 1.03% |
| Operation | Cycles (M) | % | MACs (M) | MACs/Cyc |
|---|---|---|---|---|
| Pre-Attention Norm | 2.2 | 6.2% | 0.4 | 0.17 |
| Multi-Head Attention | 18.6 | 51.9% | 120.4 | 6.46 |
| Q/K/V Projections | 5.7 | 15.9% | 44.2 | 7.76 |
| Permute/Reshape | 0.1 | 0.1% | — | — |
| QK T Attention Scores | 3.4 | 9.4% | 30.7 | 9.08 |
| Softmax Normalization | 4.5 | 12.6% | — | — |